Adding four facts from a company's latest 10-K raises gross Sharpe by 0.60 in a daily news-sentiment book already earning 2.55. That gain is the paper's finding worth paying attention to. Whether trading costs leave any of it intact remains unanswered; the paper does not attempt that test.
A frozen sentiment LLM gets filing context
Iacovides, Hannibal, Zhou and Mandic use FinDPO, a finance-tuned causal LLM that scores news sentiment. They freeze the model and call their setup FinLLM-KG. There is no retraining. The change happens in the prompt: articles arrive with company background instead of being read alone. That background comes from FinReflectKG, a published graph of (head, relation, tail) facts extracted from 10-K filings for more than 700 companies across fiscal years 2014 to 2024.
Each fact carries its EDGAR filing date, and the model can use only the latest 10-K filed strictly before the article. The premise is straightforward. When the FAA rejects Boeing's wiring proposal, knowing who regulates what and which product lines matter might help the model rank stocks better.
The authors supply background in two forms. Claude Haiku 4.5 writes a company profile under five fixed headings (Business, Financials, Momentum, Strategy, Key Risks). The cap is 500 words in the first year and 700 thereafter; each year's profile updates the previous version. Alternatively, the model receives the k = 4 triples closest to the article, measured by cosine similarity between all-MiniLM-L12-v2 embeddings. The prompt instructs the model to favor current news when background conflicts with it. A logit-to-score converter from the FinDPO work maps the label to a continuous score in [-1, 1].
The daily long-short test covers S&P 100 names from February 2015 to June 2021, giving 1,672 return days. A BERT-base-NER filter reduces roughly 200,000 articles to about 160,000, retaining only articles where the target company is tagged with over 98% confidence. Names are ranked each day. The top 35% are bought and the bottom 35% sold short, with equal weights, entry at the next open and exit at the following open.
The article-only baseline returns 275.04% cumulatively and 40.39% annualized, with Sharpe 2.55. Profiles produce 325.06%, 47.98% and 3.07. Triples produce 329.54%, 48.69% and 3.15. Sortino rises from 4.12 to 4.60 and 4.96; Calmar rises from 4.38 to 6.07 and 6.12. The reported SPY betas range from -0.04 to -0.06, which the authors take as evidence that the signal drives returns. The S&P 500 line (12.77%, Sharpe 0.72) therefore adds little to the comparison. The article-only baseline is the relevant hurdle.
What passed the filing gate?
The temporal rule is precise: use the filing date, require it to fall strictly before the article, and leave the score undefined when no filing exists. The paper also specifies k, the encoder and similarity metric, the profile headings and word caps, the NER threshold, the 35% cutoff and t+1 open execution.
Other choices are harder to recover. Retrieval happens "after removing generic relations," but we did not find the excluded relation list. Nor did we find the profile-writing or scoring prompts. The universe consists of S&P 100 constituents, filtered daily for news availability. The authors describe it as "point-in-time and dynamic," referring to the availability of news and market data. We could not establish from the text whether membership itself is historical. Using today's S&P 100 membership throughout would introduce survivorship into the 1,672-day test.
Coverage creates another uncertainty. The knowledge-enhanced score is "left undefined" when no prior filing exists, so the enhanced books may trade a different article set from the baseline. We did not find coverage counts. A replication should rerun the baseline on exactly the articles scored by the triples version. Only then does the 0.60 Sharpe gap compare like with like.
The 0.60 Sharpe gain from 10-K triples has little room for costs
Triples beat the baseline by 8.30 annualized points. Across roughly 252 trading days, that amounts to about 3.3 bps a day, gross. Table 1 explicitly reports results "at no transaction cost," and we did not find turnover. Both books rebalance daily using the same 35/35 structure, so both incur costs. Even slightly higher churn in the context-enhanced ranks could consume the entire gain through a few bps a day of additional expense.
The authors conclude that compact triples deliver the strongest gains, ahead of profiles. Table 1 supports that claim by narrow margins. Top-k leads across all five metrics, with an advantage of 0.08 in Sharpe and no more than 0.36 in Sortino. We found no confidence interval around this single backtest path. The paper calls k of 4 the "default" and shows no sweep.
We would not rank the variants on margins this small.
Look-ahead remains an open concern. The graph, built through an LLM critic-corrector loop, comes from a 2025 paper. Claude Haiku 4.5, which writes the profiles, is a current-generation LLM. The filing-date rule controls which facts reach the prompt. The profile writer's training may nevertheless contain knowledge of how 2016 to 2021 played out. Frozen FinDPO raises the same concern across every condition, including the baseline. The additional sources of leakage in the enhanced books are the profile writer and graph builder.
A Boeing 737 MAX story moves from -0.866 to +0.790 with four 10-K facts
The paper's showcase is a Reuters article dated 11 March 2020. The FAA had rejected Boeing's proposal to leave 737 MAX wiring bundles in place; those bundles may sit too close together at more than a dozen locations. Given the company prompt without triples, the model assigns -0.866. The four retrieved triples have exact Shapley values of -0.104, +0.514, +0.667 and +0.579. They sum to +1.656, taking the final score to +0.790.
The attribution arithmetic is clean. The method exposes how facts such as "FAA regulates 737 MAX return to service" turn a regulatory setback into a +0.790 score. Although the prompt tells the model to favor current news in a conflict, background wins here by 1.656. Attribution reveals the reversal; it cannot establish that the reversal was correct. With one example shown, we also cannot judge its frequency. The -0.866 is the score from the company-anchored prompt. The paper does not report the article-only FinDPO score for this story.
Our backtest stops at the missing context
We could not backtest the 10-K context method.
FinReflectKG is published, but our data contain neither that graph nor raw 10-K narrative text. Structured financials lack the relationship facts retrieved by the method. Our news history begins around 2020, after most of the February 2015 to June 2021 test window.
To move from interested to convinced, we need turnover for each book, a baseline rerun on the triples version's article set, and a net Sharpe gap that remains positive. Until then, the move from 2.55 to 3.15 is a gross ranking improvement along one backtest path.